01
Click endurance
29k
saved
32.2%
less token use
91k baseline to 62k with Kairn.
Local-first v0.2 public beta
Local repository evidence for the coding agents you already use.
Install once. Kairn points supported agents toward the right source and reduces unnecessary searching, repeated reads, and tokens — without uploading repository evidence or agent trajectories to Kairn.
$
npm install -g usekairn@betaRegistry-verified: 0.2.0-beta.2. Then run `kairn enable` inside a repository.
Fully local core. No Kairn account, hosted backend, or Kairn API key.
Less token waste. Savings count only after required-source and quality checks pass.
Agent-ready. Use local hooks or MCP with supported coding agents.
What it does
Turn the current checkout into useful evidence.
Kairn is not generic agent memory. It is a local evidence layer focused on the repository state in front of you now.
01
Map the current checkout without uploading repository contents.
02
Use paths, symbols, failures, git state, and session history to find likely production source.
03
Send compact evidence or a path hint instead of dumping the repository into the prompt.
04
Avoid reopening unchanged context across a long coding session.
05
Stay quiet or ask for missing evidence when extra context would be guesswork.
How it works
Repository signals in. Small, evidence-backed scope out.
Kairn v0.2 combines a cheap local repository map with bounded evidence retrieval and deterministic safety decisions. A calibrated local ranker is planned for v0.3 and will not be marketed as shipped before it clears its holdout gates.
01
Map
Inventory paths, symbols, source roles, and direct repository relationships.
02
Gather
Start with exact paths, failures, symbols, lexical matches, git state, and session evidence.
03
Decide
Rank a bounded candidate set and suppress low-value or repeated context.
04
Deliver
Give the agent a compact path hint or source payload through local hooks or MCP.
Why it matters
Local evidence and verified token efficiency.
Right source sooner
Kairn prioritizes current-checkout evidence so the agent can begin with likely production files instead of broad exploration.
Verified token efficiency
The goal is fewer searches, repeated reads, and repository-content tokens. A token win only counts when the task still passes.
Local by default
Core indexing, ranking, and session state run on your machine without a Kairn account, API key, or hosted data plane.
Agent-independent
Kairn works beside supported tools through local hooks and MCP instead of requiring control over the agent's internal loop.
Open distribution
The published v0.2 beta package is Apache-2.0 licensed. A public GitHub source release is planned for v0.3; the repository remains private today.
Legacy pre-ranker evidence
The current public numbers are historical, quality-gated beta evidence from before the calibrated ranker and frozen comparison protocol. They remain visible as a baseline, not as proof of the planned v0.3 system.
Evidence snapshot
Observed token reductions, after quality passes.
Evidence sets
8+
historical pre-ranker sets
Median reduction
32.2%
observed tokens, clean rows
Official tasks passed
20/25
historical official campaigns
Endurance suites
4
historical long-session evidence
02
Node SemVer endurance
11k
saved
16.4%
less token use
70k baseline to 58k with Kairn.
03
OpenLibrary SWE-Pro
12k
saved
54.2%
less token use
22k baseline to 10k with Kairn.
These are observed model-token comparisons, not estimates from live telemetry. Savings vary by task and count only on pass/pass rows; the detailed page includes regressions and quality-rescue rows separately.
Supported vs. validated
One local layer, several agent paths.
`kairn enable` detects supported clients. Support means setup is available; validation means that path has passed a published campaign.
Codex CLI
Validated
The most-tested beta path for source routing and long sessions.
Local MCP
Validated transport
Real MCP calls have been exercised in existing benchmark rows.
Claude Code
Supported setup
Local MCP configuration is supported; comparative efficacy validation is pending.
Gemini CLI
Supported setup
Local MCP configuration is supported; comparative efficacy validation is pending.
VS Code / Copilot
Supported setup
Local MCP configuration is supported; comparative efficacy validation is pending.
A focused product
What Kairn is not building.
Kairn owns precise, local, current-repository evidence. It does not need to become an agent platform to do that well.
Not reasoning-trace capture.
Not model routing.
Not hosted collective memory.
Not trajectory dashboards.
Not a replacement coding agent.
Privacy
Local by default.
Repository evidence, prompts, diffs, paths, and agent trajectories are not sent to Kairn. Optional aggregate telemetry is off until you enable it.
Read the privacy boundaryOpen distribution
Apache-2.0 beta package, today.
The published v0.2 beta package is licensed under Apache-2.0. Kairn's GitHub source repository remains private today; its public source release and stable v0.3 distribution are planned.
Public beta